Received 31 March 2024; revised 1 July 2024; accepted 15 July 2024.

Date of publication 10 September 2024; date of current version 29 November 2024.

This article has supplementary downloadable material available at

https://doi.org/10.1109/TVCG.2024.3456340, provided by the authors.

Digital Object Identier no. 10.1109/TVCG.2024.3456340

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

973

See https://www.ieee.org/publications/rights/index.html for more information.

DG Comics: Semi-Automatically Authoring

Graph Comics for Dynamic Graphs

Joohee Kim , Hyunwook Lee , Duc M. Nguyen, Minjeong Shin , Bum Chul Kwon ,

Sungahn Ko

∗

, and Niklas Elmqvist

Supporter Top 30%

Main Character

France 0.33 0.33 0.33

Ecuador 0.06 0.06 0.06

El Salvador 0.06 0.06 0.06

Japan 0.32 0.32 0.32

Mexico 0.14 0.14 0.14

Iraq 0.06 0.06 0.06

Finland 0.06 0.06 0.06

Sweden 0.1 0.1 0.1

Yugoslavia 0.19 0.19 0.19

Iran 0.06 0.06 0.06

Australia 0.14 0.14 0.14

1 2

Search node

Germany

Poland

Romania

Estonia

Latvia

In August 1939, Germany

and the Soviet Union agreed

to not attack each other and

divide the countries

between them secretly.

Germany, Japan, and Italy

formally became a military alliance,

called the

Axis powers

United States of America

France

Germany

Spain

Switzerland

Romania

Hungary

United States of America

United Kingdom

France

Germany

Switzerland

Canada

Egypt

United States of America

United Kingdom

Netherlands

France

Sweden

Switzerland

Czechoslovakia

United States of America

United Kingdom

Italy

Japan

Saudi Arabia

Iraq

Egypt

Syria

Lebanon

Jordan

German Federal Republic

Sankey Node Min. Max Avg Class

Ribbentrop Pact

1941

Level 1.0

Supporter Top 30% Highlighter

Main Character

France 0.33 0.33 0.33

Ecuador 0.06 0.06 0.06

El Salvador 0.06 0.06 0.06

Japan 0.32 0.32 0.32

Mexico 0.14 0.14 0.14

Iraq 0.06 0.06 0.06

Finland 0.06 0.06 0.06

Sweden 0.1 0.1 0.1

Yugoslavia 0.19 0.19 0.19

Iran 0.06 0.06 0.06

Australia 0.14 0.14 0.14

1 2

Supporting Character

Australia

Iran

Finland

Yugosla

Brazil

Chile

Bulgaria

Russia

Turkey

Canada

South Africa

Egypt

Venezuela

1 2

France

Search node

Germany

Russia

Poland

Romania

Estonia

Latvia

Finland

In August 1939, Germany

and the Soviet Union agreed

to not attack each other and

divide the countries

between them secretly.

Meanwhile, Germany, Japan, and Italy

formally became a military alliance, called

theAxis powers

United States of America

France

Germany

Spain

Switzerland

Romania

Hungary

United States of America

United Kingdom

France

Germany

Switzerland

Canada

Egypt

United States of America

United Kingdom

Netherlands

France

Sweden

Switzerland

Czechoslovakia

United States of America

United Kingdom

Italy

Japan

Saudi Arabia

Iraq

Egypt

Syria

Lebanon

Jordan

German Federal Republic

Sankey Node Min. Max Avg Class

Sankey Node

Germany

Poland

United Kingdom

France

Germany

World War II was sparked by the

Nazi German invasion of Poland in 1939.



Britain and France declared war on

Germany on 3 September 1939

.

Phoney

 W

Molotov–

Ribbentrop Pact

 However, in June 1941, it was

terminated due to the German

invasion of the Soviet Union.

Graph Comic Basic Filled Fixed

Timeline ADD REPL

1941

Level 1.0

Supporter Top 30% Highlighter Top 10%

Main Character

France 0.33 0.33 0.33

Ecuador 0.06 0.06 0.06

El Salvador 0.06 0.06 0.06

Japan 0.32 0.32 0.32

Mexico 0.14 0.14 0.14

Iraq 0.06 0.06 0.06

Finland 0.06 0.06 0.06

Sweden 0.1 0.1 0.1

Yugoslavia 0.19 0.19 0.19

Iran 0.06 0.06 0.06

Australia 0.14 0.14 0.14

1 2

Supporting Character

EDIT

Australia 38.56

Iran 38.2

Finland 37.99

Yugoslavia 23.55

Brazil 23.52

Chile 20.27

Bulgaria 19.04

Russia 15.67

Turkey 10.86

Canada 9.67

South Africa 8.81

Egypt 7.83

Venezuela 6.31

1 2

France

Summary Node Attribute Community

GENERATE

Search node

Germany

Soviet Union

Poland

Romania

Estonia

Latvia

Finland

In August 1939, Germany

and the Soviet Union agreed

to not attack each other and

divide the countries

between them secretly.

Meanwhile, Germany, Japan, and Italy

formally became a military alliance, called

the

Axis powers

United States of America

France

Germany

Spain

Switzerland

Romania

Hungary

United States of America

United Kingdom

France

Germany

Switzerland

Canada

Egypt

United States of America

United Kingdom

Netherlands

France

Sweden

Switzerland

Czechoslovakia

United States of America

United Kingdom

Italy

Japan

Saudi Arabia

Iraq

Egypt

Syria

Lebanon

Jordan

German Federal Republic

Sankey Node Min. Max Avg Class

Sankey Node Weight

Germany

Poland

United Kingdom

France

Germany

World War II was sparked by the

Nazi German invasion of Poland in 1939.

Britain and France declared war on

Germany on 3 September 1939.

Phoney

 War

Molotov–

Ribbentrop Pact

 However, in June 1941, it was

terminated due to the German

invasion of the Soviet Union.

United States of America

United Kingdom

France

Germany

Sweden

Poland

Austria

Czechoslovakia

France

Italy

China

Japan

Hawaii

Yugoslavia

Singapore

Ethiopia

Cambodia

Hungary

Solomon Islands

Hong Kong

British Somaliland

Greece

Romania

Bulga

ria

United States of America

United Kingdom

Netherlands

France

Sweden

Switzerland

Czechoslovakia

United States of America

United Kingdom

Italy

Japan

Saudi Arabia

Iraq

Egypt

Syria

Lebanon

Jordan

German Federal Republic

font

1931

1932

1933

1934

United Kingdom

China

United Kingdom

China

United Kingdom

China

The Big Three

—the United Kindom, the United

States and the Soviet Union

—, China and France

officially formed

the United Nations to oppose

Italy

China

Hawaii

Yugoslavia

Singapore

Ethiopia

Cambodia

Solomon Islands

Hong Kong

British Somaliland

Bulgaria

Germany

Soviet Union

 However, i

n June 1941, it was

terminated due to the German

invasion of the Soviet Union.

The Allies

The Axis powers rapidly expanded their influence in

Europe, Africa, and Asia.

The attack on Pearl Harbor led

the U.S. to formally enter WWII

24. 4. 1. 오후 4:07

Graph Comic

Timeline ADD

United States of America

United Kingdom

France

Soviet Union

China

Germany

Battle of Midway

Battle of Stalingrad

On May 8, 1945, World War II in Europe ended.

Japan surrendered to the Allies on September 2.

The End of

World War II

Japan

Fig. 1: World War II. Graph comic on the history of World War II, covering the causes and signiﬁcant events, such as the formation

of alliances and major battles, created using DG Comics. Note that while the under lying dynamic graph comic was generated

automatically, the DG Comics tool provides functionality for the designer to manually move nodes, add visual highlighting, and insert

the geographic map backgrounds. We used contemporary national ﬂags for easier recognition wherever feasible.

Abstract—Comics are an effective method for sequential data-driven storytelling, especially for dynamic graphs—graphs whose

vertices and edges change over time. However, manually creating such comics is currently time-consuming, complex, and error-prone.

In this paper, we propose DG COMICS, a novel comic authoring tool for dynamic graphs that allows users to semi-automatically build

and annotate comics. The tool uses a newly developed hierarchical clustering algorithm to segment consecutive snapshots of dynamic

graphs while preserving their chronological order. It also presents rich information on both individuals and communities extracted from

dynamic graphs in multiple views, where users can explore dynamic graphs and choose what to tell in comics. For evaluation, we

provide an example and report the results of a user study and an expert review.

Index Terms—Data-driven storytelling, narrative visualization, dynamic graphs, graph comics

1 INTRODUCTION

Dynamic systems are prevalent in both nature and society. Catalysts

facilitate chemical reactions, while species interact throughout evolu-

∗

Corresponding author

• Joohee Kim, Hyunwook Lee, Duc M. Nguyen, and Sungahn Ko are with

UNIST (Ulsan National Institute of Science and Technology).

E-mail: {joohee, gusdnr0916, ducnm, sako}@unist.ac.kr

•

Bum Chul Kwon is with IBM Research. E-mail: bumchul.kwon@us.ibm.com

• Niklas Elmqvist is with Aarhus University. E-mail: elm@cs.au.dk

tion. Scientists collaborate with colleagues and students across various

stages of their careers. Social relationships form, evolve, and dissolve

as individuals make friends, have children, and experience rifts or pass

away. Modeling these phenomena as dynamic graphs, where nodes rep-

resent entities and links represent their evolving relationships over time,

is a valuable tool for understanding such systems. However, due to

inherent complexity and scale, it is challenging to communicate stories

extracted from dynamic graphs succinctly and accurately. In previous

studies, graph comics, a comic-based storytelling medium that consists

of graph visualizations, is found to be effective in narrating stories

involving dynamic graphs [4, 65,86]. Despite the potential beneﬁts, it

is time-consuming for users to create such graph comics based on their

data, especially given the lack of dedicated authoring tools.

We address this gap by proposing an approach to automatically

generate graph comics from hierarchical clustering that preserves the

temporal causality (thus causality-preserving) of the events. To validate

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

974

the approach, we implement an interactive tool called DG COMICS:

Storytelling Automation: To combat the complexity of dynamic

graphs with many relationships spanning a long period of time,

DG Comics automates the clustering of temporal events into seg-

ments using causality-preserving hierarchical aggregation [24].

Storytelling Agency: To facilitate the creative and functional

agency of the analyst authoring the story, DG Comics yields con-

trol of the aggregation level, the main and supporting characters

to display, as well as style and formatting choices to the analyst.

DG Comics offers a causality-preserving clustering of graph snap-

shots in a dendrogram, which users can leverage to initiate graph comic

generation. From a chosen aggregation level, which determines the

number of panels, DG Comics automatically creates a comic template

with a computed layout. Each panel features the main character(s)

undergoing the most signiﬁcant changes and their relationships. These

changes are depicted with distinct visual representations and captions

generated using a template-based approach. Users can further develop

the template using various editing functions and build original stories

through interaction with additional views and tables. We veriﬁed DG

Comics’s usefulness and versatility through (1) an example using a

VIS coauthorship network, (2) a user study involving 13 university

participants designing graph comics with an international trade network

dataset, and (3) interviews with six experts from various domains.

The contributions of this work include (1) robust design requirements

derived from challenges identiﬁed in prior research; (2) development of

a causality-preserving clustering technique with enhanced graph simi-

larity computation as the backbone of automation; (3) implementation

of a novel interactive comic authoring tool; and (4) results from an

example scenario, a user study, and expert reviews.

ELATED WORK

Our work lies at the intersection of graph visualization [75], data-driven

storytelling [59], and data comics [86]. Below we review the literature

in all these areas and discuss how our work supersedes prior art.

2.1 Graph Visualization

A graph or a network consists of nodes and links, with nodes rep-

resenting entities (e.g., people, organizations) and links representing

relationships (e.g., friendships, alliances). Graphs are widely used in

domains [75] such as transportation, biology, communication, business,

and security. As their utility increases, so does their size. Combin-

ing networks with data like maps (geospatial graphs), time (dynamic

graphs) [25], or text adds complexity. Conventional graph visualiza-

tions include two methods: node-link diagrams [28], which use lines

to connect nodes and visual channels (e.g., color, size, line thickness)

to distinguish attributes, and adjacency matrices [12, 30, 31], which

organize nodes in a grid and display connections at intersecting cells.

Prior art [51, 55, 75] provides detailed descriptions.

As graphs grow, visualizations become complex and cluttered. Net-

workNarratives [46] provides semi-automated guided data tours to

facilitate the navigation of complex networks. Symbolic representa-

tions are one way to overcome this issue by aggregating or hiding nodes.

Dunne and Shneiderman [23] propose a simpliﬁcation technique that

uses fan, connector, and clique motifs to save space and improve un-

derstanding of large graphs. To prevent misunderstanding and address

the complexity of large graphs, some research [30, 31] combines node-

link diagrams and adjacency matrices, enabling efﬁcient exploration

of networks. Yoghourdjian et al. [83] introduce graph thumbnails for

high-level structure visualization, allowing easy identiﬁcation, com-

parison, and overview of multiple large graphs. Ghani et al. [28] use

dynamic insets to show off-screen node neighborhoods, while May

et al. [49] improve off-screen awareness by providing graph neighbor

information. Visualizing groups or clusters in graphs is another area

of research. Saket et al. [63] propose a taxonomy for graph groups,

enumerating tasks such as group-only, group-node, group-link, and

group-network. Vehlow et al. [74] survey techniques for visually pre-

senting graph groups, categorizing them into visual node attributes,

juxtaposed, superimposed, and embedded methods.

2.2 Visualizing Dynamic Graphs

Dynamic graphs are graphs whose relations among entities change over

time. Such changes bring challenges in tasks, visualization, and evalua-

tion, prompting signiﬁcant research efforts. Ahn et al. [1] categorize

temporal features of dynamic graphs by the rate and shape of changes

and individual events. Beck et al. [11] survey dynamic graph visualiza-

tion techniques, focusing on presentation methods such as animated di-

agrams and static timeline-based charts. Analyses of temporal features

from individual node/link, group, and network perspectives inspired

how DG Comics presents information on dynamic graphs. Numerous

techniques have been developed for visualizing dynamic graphs. Elzen

et al. [71] visualize graph evolution with sequence views, and Burch

et al. [16] propose pixel-oriented visualizations. GeneaQuilts [13] use

a hybrid adjacency matrix and node-link representation to visualize

family trees over time. Bach et al. [3, 5] employ 3D matrix cubes and

small multiples of adjacency matrices.

To reduce the complexity of dynamic graphs, computational methods

like spectral graph wavelets [22] and diachronic node embedding [81]

hierarchically aggregate [24] graph snapshots based on graph struc-

tures [2] or attributes [29]. The computation results effectively visualize

the changes by time (e.g., small multiples [3]). For example, Elzen et

al. [72] present a novel visual analytics pipeline that allows users to

track graph changes with points. The pipeline consists of democrati-

zation, vectorization and normalization, dimensional reduction, and

visualization, transforming graph snapshots into points. Cakmak et

al. [18] propose multi-scale snapshot visualization with graph2vec [53],

creating temporal summaries of dynamics graphs. We refer to [80] for

an extensive survey result on graph learning algorithms.

2.3 Graph Comics and Authoring Tools

Comics [50] are a storytelling genre that presents stories with combina-

tions of illustrations, text, and annotations on various layouts.Segel and

Heer propose using comics for data-driven storytelling in their seminal

work on narrative visualization [65]. Data comics, an emerging medium

of data-driven narrative visualization, leverage the visual language of

comics, including layouts, characters, and captions [6,86]. Research has

focused on identifying characteristics of data comics [70], developing

authoring tools [20, 86], and speciﬁc applications such as visualization

education [76] and user study reports [77]. Bach et al. [4] conduct a

design study on storytelling with dynamic networks, proposing design

factors: visual representation of graph elements and changes, tempo-

rality of changes, element identity, cast of characters, level of detail,

overview and detail, and representation of multivariate networks. They

also deﬁne a design space for data comics by analyzing common pat-

terns in existing storytelling media (e.g., infographics, data videos) [7].

This design space has two dimensions: content relation patterns (e.g.,

narrative, temporal, faceting, visual encoding, granular, spatial patterns)

and panel layout (e.g., linear, tiled, parallel, grid). Wang et al. [79] con-

duct controlled and in-the-wild user studies to investigate the beneﬁts

of data comics compared to infographics. Their experiments reveal that

data comics improve understanding and recall of information, and are

preferred for enjoyment, focus, and engagement.

Several data comics authoring tools have been proposed for various

computing environments (e.g., tablets, computational notebooks, cod-

ing environments). DataToon [40] is the ﬁrst effort to help users design

comics on dynamic graphs with pen and touch interaction. Computa-

tional notebooks are a new tool for analyzing, visualizing, and sharing

datasets. However, their results, which combine programming code,

notes, and analysis, often struggle to communicate with the audience.

To address this, Kang et al. [38] propose ToonNote, an extension that

converts notebooks into data comics. CodeToon [69], similar to Toon-

Note in motivation, focuses more on coding environments, facilitating

code-aligned storytelling and automated comic generation. Most data

comics are static to guide readers through a speciﬁc ﬂow and layout.

Wang et al. [78], posing questions on the interactivity of data comics,

formalize operations for interactive comics (e.g., content highlighting,

panel addition/removal). They also present a lightweight scripting

approach with six goals: navigation, details on demand, changing

perspective, branching, pause and reveal, and input data.

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

975

kim ET AL.: DG COmiCS: SEmi-AUTOmATiCALLY AUTHORiNG GRAPH COmiCS FOR DYNAmiC GRAPHS

2.4 Comparison to Prior Art

Our proposed work in this paper uses comics to visualize dynamic

graphs changing over time, and our approach is novel over the literature:

•

Building on Segel and Heer [65]’s taxonomy, Zhao et al. [86] ﬁrst

proposed data comics for narrative visualization, but our approach

goes beyond their work by applying the idea to graphs.

•

While Bach et al. [4] apply comics to graphs, our work proposes

automatic graph comic generation as well as an interactive editor.

•

Authoring tools for data comics exist [38, 40, 69, 86], but ours is

unique to dynamic graphs and semi-automatically builds comics.

•

Prior work visualizes temporal summaries of dynamic graphs [3,

18, 85], but we apply the idea to automating data comics.

HALLENGES AND DESIGN REQUIREMENTS

The aim of this paper is to develop an authoring tool for data comics that

automatically generates a narrative in comic form and enables users

to efﬁciently reorganize them into a coherent storyline. We accom-

plish this by adhering to the two design principles outlined in Sec. 1:

storytelling automation versus storytelling agency. These principles

guide authors in managing scale and complexity while preserving their

creative and expressive vision for the data-driven story.

To achieve these goals, we reviewed existing fundamental design

principles formalized for storytelling [7, 33, 42, 45, 65, 68] and data

comics [4, 6, 7, 84, 86]. We also based our work on prior research in

dynamic graphs regarding visual analysis [37, 51, 74, 75]), tasks [1, 11,

25], and computational methods [22, 72, 81], as well as data comics

authoring tools [38, 40, 69, 78]. Based on the design principles and

literature survey, we derive several challenges, as listed below.

C1–Size and Complexity. Creating comics from dynamic graph

data demands identifying key narrative elements, such as pivotal nodes

or events, for storytelling. This task ranges from emphasizing criti-

cal nodes to elaborating on connections for speciﬁc events. Viewing

dynamic graphs through various lenses, like individual nodes or node

communities, can reveal these elements. However, analyzing dynamic

graphs across multiple perspectives and timeframes is challenging due

to the size and complexity of general dynamic networks.

C2–Identifying Characters. Because data comics are a genre

of comics, they follow common design patterns in comics [50]. An

important pattern is the existence of main and supporting charac-

ters [6, 42, 65, 70, 86]. Main characters drive the story while supporting

characters play key roles in the narrative. Identifying which nodes serve

as main or supporting characters poses a challenge, given the multitude

of nodes with varying and evolving characteristics.

C3–Changes Over Time. The single most important task in a dy-

namic graph visualization is to show changes in the graph over time [1],

particularly for the main and supporting characters, or their relation-

ships. The challenge lies in detecting the changes, presenting them,

and annotating the story. Detecting temporal changes requires a deep

understanding of the data, and is difﬁcult to do manually, especially for

large graphs. Visualizing such changes is also not a trivial problem. Fi-

nally, annotations are essential for explaining the context or conveying

additional insights.

C4–The Language of Comics. Comics employ a distinctive

narrative structure, visual language, and temporal sequencing to engage

readers. For end-users unfamiliar with these conventions, effectively

integrating them into dynamic graph comics can be daunting. Ensuring

these elements are used effectively to convey complex information in

an accessible and compelling manner requires a deep understanding of

comic artistry and narrative technique.

Design Requirements. Below we present design requirements

(R1–R5) for our proposed data comics authoring tool that can help

users overcome these challenges (C1–C4):

Allow users to choose temporal granularities for the data comics

that reﬂect their desired levels of detail (C1).

Enable multi-perspective storytelling (individual, community-

based, metric-based, etc) (C1).

Help users ﬁnd main characters and supporters for initiating a

narrative (C2).

Support users to detect and present temporal changes of dy-

namic graphs with appropriate annotations (C3).

Provide comics authoring mechanisms to assist users in follow-

ing genre conventions (C4).

RAPH COMIC GENERATION TECHNIQUES

We introduce techniques for streamlining the generation of graph

comics. We develop a hierarchical clustering algorithm that groups

snapshots into different temporal granularities based on their similarity.

This clustering produces a dendrogram, enabling users to automatically

create comics by selecting the number of panels.

4.1 Visualizing Dynamic Graphs as Comics

A graph snapshot is a static graph representing the state of a dynamic

graph at a speciﬁc point in time. We can consider a dynamic graph

as a collection of graph snapshots: snapshots organized in a temporal

sequence, where a snapshot is generated whenever there is a change in

the dynamic graph, such as the addition or deletion of a node or link.

Such a dynamic graph can be naïvely visualized as a data comic by

rendering one comic panel per graph snapshot and then visualizing the

speciﬁc snapshot inside its panel as a node-link diagram or adjacency

matrix. We provide continuity between adjacent panels by freezing the

position of nodes from one panel to the next and emphasizing changes—

node and link additions and deletions—using visual highlighting, such

as colors, callouts, and visual effects. The result is a graph comic [4]: a

sequence of panels showing a graph changing over time.

Unfortunately, this naïve approach is not practical for graphs span-

ning long time periods or involving many changes because the resulting

comic will have a prohibitively large number of panels. Furthermore,

large graphs with many nodes and links will yield panels that are so

cluttered that individual changes are hard to spot. Below we describe

practical approaches to address these shortcomings:

•

Long times: We present a causality-preserving temporal clus-

tering algorithm that combines multiple adjacent snapshots into

clusters to support balancing the number of panels and the amount

of changes in each panel based on the user’s needs; and

•

Large graphs: We propose graph ﬁltering mechanisms based

on the concept of main and supporting characters so that large

graphs can be reduced to more manageable subgraphs.

4.2 Causality-Preserving Temporal Clustering

We present an algorithm that hierarchically groups graph snapshots in

a dynamic graph based on the similarity between adjacent snapshots

while preserving their causal (temporal) order. Instead of showing every

graph snapshot, the graph comic can show snapshot groups consisting

of the union of several temporally adjacent snapshots. The goal is to

facilitate users precisely adapting the number of panels to show the

resulting comic, from a single panel representing all changes (i.e., a

snapshot group representing all snapshots) to a panel for every individ-

ual change in the graph. We consider four additional requirements for

our algorithm: it must take into account the node and link labels, which

are crucial for the graph data; it should preserve temporal continuity

and consistency, meaning it should provide the same similarity for the

same input; it should factor in multiple numerical attributes for both

nodes and links; and it should not be computationally heavy.

4.2.1 Adjacency-based Hierarchical Clustering

We employ a variant of agglomerative clustering [24] where, instead

of comparing the distance from each graph snapshot to every other

snapshot, we only compare the distances between snapshots that are ad-

jacent in time [85]. In other words, given a dynamic graph

consisting

of a sequence of graph snapshots

for each time

t ∈ [0 ...T ]

for the

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

976

Fig. 2: Comparison of graph distance metrics for two graphs.

and

t+1

with different labels and attributes with (a) Weisfeiler-Lehman

kernel-based distance, (b) deep learning-based similarity (or distance),

graph edit distance with weighted Jaccard similarity (our choice). The

thickness of the line is proportional to the node or link attributes.

time period

, only graph snapshots at

t − 1

and

t +1

are compared (i.e.,

snapshots just before or after the current snapshot). This means that the

temporal order is preserved when graph snapshots are aggregated into

snapshot groups, thus maintaining causality in the resulting comic.

Conceptually, a snapshot group is regarded the same as a snapshot:

it has a time span instead of a point in time representing the ﬁrst and

last time stamps for its constituent graph snapshots, and its graph is the

union of those constituent graphs. Importantly, the distance metric is

deﬁned similarly for both a snapshot and a snapshot group.

To build a cluster hierarchy of a sequence of graph snapshots (and

snapshot groups), we must provide a distance metric

D(G

, G

)

that

accepts two adjacent graph snapshots (or groups)

and

. We

agglomerate the dynamic graph sequence by progressively selecting the

two adjacent snapshots with the smallest distance and replacing them

with a snapshot group combining them. This means that the number of

snapshots in the sequence will monotonically decrease until a single

snapshot group representing all of the original graph snapshots remains.

4.2.2 Graph Edit Distance with Weighted Jaccard Similarity

Several graph similarity measures exist in the literature. Basic graph

topology measures for unweighted, unlabeled, and undirected graphs

are not good metrics of similarity, as real-world dynamic graphs often

have both node and link attributes that factor into similarity. Next, we

review existing work on graph distance measure metrics and propose

our new distance measure metric used in this work.

We ﬁnd four types of metrics: graph edit distance (GED), set simi-

larity, kernel-based measures, and deep learning-based measures. GED

calculates a weighted summation of predeﬁned graph edit operators,

such as insertion, deletion, and substitution of a node or link. Although

simple and effective, GED requires human effort to determine each

edit operator and the corresponding cost. Furthermore, existing GED

methods deﬁne substitution as a link or node replacement, which is not

suitable for real-world graphs [26]. For example, as shown in Fig. 2

(c), it cannot measure the attribute change of node D.

Set similarity can be used as a graph similarity metric based on the

number of common nodes (or links) in two graphs [41, 73]. The main

advantage of this method is that it has linear computational complexity

and is suitable for real-world applications because of its label awareness.

However, similar to GEDs, it cannot measure attribute changes.

Kernel-based methods using graph topology are another line of

research for graph similarity computation. Examples include path-

based kernels [27, 39], subtree pattern-based graph kernels [48,58], and

Weisfeiler-Lehman optimal assignment kernels [43]. While effective

in measuring the overall similarity over the entire graphs, they often

omit node attributes (R2) and cannot be used for varying aggregation

levels (R2). In addition, as shown in Figure 2 (a), it may lose the label

information (e.g., treating link (A, E) and (E, F) as the same one).

Deep learning-based methods [47, 82] have been developed to over-

come the heavy computational costs in measuring graph similarity.

However, the explainability of deep learning methods remains an open

problem, so we do not use them in this work, as our graph comics need

multi-perspective storytelling (R2) with solid reasoning (R3–R4).

After reviewing existing metrics and their characteristics, we ﬁnd

that traditional methods, including GED and set similarity, are the

most appropriate metrics for our purpose. However, they cannot be

directly used for our work, as they are not able to measure attribute

changes. As such, we develop an enhanced method with GED and

weighted Jaccard similarity. Our method calculates vector-form Jaccard

similarity [62] for each common node or link. For example, as shown

in Fig. 2, we ﬁnd common elements with labels (e.g.,

(B, E)

) and

then calculate vector-form Jaccard similarity for the attributes of each

common element. As a result, we can obtain the degree of changes per

link or node (e.g., change of attributes for node D), as in Fig. 2 (d).

HE DG COMICS SYSTEM

We design DG Comics to meet the challenges and satisfy the require-

ments from Sec. 3. Our approach outlined in Sec. 4 efﬁciently manages

complexity (R1) for automatic graph comic generation. The algorithm

output, visually represented as a dendrogram (Sec. 5.1), allows users

to interactively choose an aggregation level for story fragments, which

can then be depicted as comic panels (Sec. 5.2). Each comic panel

represents a story fragment, and the node-link diagram in the panel

visualizes changes during that interval (R4). Users can build their own

stories by inspecting potential main and supporting characters (R3) with

graph evaluation metrics (R2), such as node centrality, node degrees,

and adjacency (Sec. 5.4), as well as individual graphs at different time

points (Sec. 5.3). DG Comics facilitates the observation of communi-

ties (Sec. 5.5) where characters are involved (R2). Finally, it enables

editing based on the language of comics (Sec. 5.2), including fonts,

motion lines, captions, and layouts (R5). DG Comics employs Next.js

for the frontend, FastAPI for the backend, and the D3 library [14] for

visualizations, including computation of node-link diagram layouts.

The source code is available at github.com/joohe-e/DGComics.git.

5.1 Summary View

Comics are made of strips of one or more panels [50, 64], each de-

picting a portion of the narrative and the sequence typically showing

temporal progression. In a graph comic, each panel contains a snapshot

of the dynamic graph sequence (either as an individual or a group) and

a caption that provides information about the snapshot. The Summary

View (Fig. 3A) facilitates the automatic generation of a series of pan-

els via a dendrogram that visualizes the hierarchical clustering result

(Sec. 4). Each leaf node of the dendrogram refers to a graph snapshot at

a speciﬁc time point. The X-axis indicates the time from the beginning

to the end of the data, and the Y-axis notes the normalized similarity be-

tween two adjacent snapshots from 0 to 1. The more similar the groups

or individual snapshots are, the lower their position is connected, as

the distance is shorter. A horizontal dashed line, called a depth slider,

partitions the dendrogram into multiple clusters, visually encoded with

different colors. The number beside the line indicates the number of

clusters formed at that depth level.

Users can select the number of panels to include in the comic strip

by moving the depth slider (R1). As users adjust the slider, the clusters

are colored differently. They can generate the comic strip by clicking

the

GENERATE

button. Each branch selected by the depth slider returns

the characters that change the most over the timespan corresponding to

the cluster (R3). For example, if a user clicks

GENERATE

after positioning

the depth slider as shown in Fig. 3A, three panels are created in the

comic view, each representing the assigned clusters (green, purple, and

orange). DG Comics offers two options for constructing subgraphs with

the ego being the resultant main character(s): a 1.0-level ego network,

which includes the ego and its 1-degree alters, and a 1.5-level ego

network which also includes the ties between the ego’s alters. Users

choose an option based on what relationship they want to present (R2).

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

977

kim ET AL.: DG COmiCS: SEmi-AUTOmATiCALLY AUTHORiNG GRAPH COmiCS FOR DYNAmiC GRAPHS

Graph Comic

Timeline ADD REPL

2011

2012

2014

2015

2016

Level 1.5

Daniel A. Keim

Supporter Top 15% Highlighter Top 5%

Main Character

Daniel A. Keim 7 24 14.44

Olivier Thonnard 6 6 6

Adam Sah 5 5 5

Jefferson

Amstutz

7 7 7

Erik Trostmann 3 3 3

Christoph

Wimmer

4 4 4

Jan Kretschmer 3 5 4

Markus Höhn 12 12 12

Aude Oliva 6 7 6.5

Kevin A. Roundy 7 7 7

Robert Kosara 1 1 1

1 2 3 4

Supporting Character

EDIT

Tobias Schreck 10

Dominik Sacha 6

Michael

Behrisch 0001

Enrico Bertini 4

Halldór

Janetzko

Dominik Jäckle 4

Oliver Deussen 3

Christopher

Collins 0001

Juri

Buchmüller

Mennatallah El-

Assady

Matthias Kraus 3

1 2

Daniel A. Keim

Tobias Schreck

Enrico Bertini

Halldór Janetzko

Michael Behrisch 0001

Dominik Sacha

Daniel A. Keim

Tobias Schreck

Michael Behrisch 0001

Daniel A. Keim

Summary Node Attribute Community

GENERATE

Search node

Sankey Node Min. Max Avg Class Sankey Node Weight

Enrico Bertini

Daniel A. Keim

Tobias Schreck

Christian Rohrdantz

Daniel A. Keim

Tobias Schreck

Dominik Sacha

Daniel A. Keim

Tobias Schreck

Dominik Sacha

Daniel A. Keim

Dominik Sacha

Tobias Schreck

Daniel A. Keim

Dominik Sacha

Tobias Schreck moved Graz University of Technology in 2015.

Dominik Sacha actively collaborated with Daniel Keim during his Ph.D.

                   Howe

ver, since the transition to an industry role,

         he st

opped publishing papers at Visualization conferences.

Mental map preservation

Community changes

D E

delete

color

size

STcolor

stroke

highlight

opacity

label

image

Lsize

cluster

Fig. 3: DG Comics overview. DG Comics offers (A) a Summary View, (B) sliders for ﬁltering and highlighting nodes, (C) a Graph Comic View, (D)

Main Character and (E) Supporting Character tables, and (F) a Timeline View. Users can switch to (G) the Node Attribute Table or (H) Community

View using the tab. It supports (M) mental map preservation by ﬁxing nodes across displays, and (O) community changes using bubble sets.

Union T1

Union T2

T2T1

T1 T2

Added

=T2-T1

Deleted

=T1-T2

Maintained

T1 T2

Selected Panel

Fig. 4: Data abstraction and presentation. Each cluster below the

depth slider (left) represents a panel (right) through a set operation of

two subcluster snapshots. The aggregate graph snapshot is computed

by the union of graphs for timespans T1 and T2.

5.2 Graph Comic View

The Graph Comic View allows users to edit graph comic templates

while managing large dynamic graphs efﬁciently. To reduce visual

clutter and enhance readability, we provide an option to keep only the

top N percent of alters based on link weights visible. Users can control

this level of visibility using supporter slider, and further highlight

speciﬁc alters with the highlighter slider. Both sliders can be adjusted

simultaneously before users begin editing (Fig. 3B).

5.2.1 Visual Design

We create a node-link diagram for each panel as the union of the

panel’s children (or the snapshot itself for leaf nodes). Each panel

visualizes the changes between two consecutive children snapshot

groups, highlighting added nodes and links with a neon effect and

deletions with dashed strokes (R4). We abstract the data for each

change using set operations: subtracting the former timespan (T1) from

the latter (T2) yields the added subgraph, while the reverse reveals the

deleted subgraph (Fig. 4). Supporting characters are deﬁned as nodes

with greater weight in their relationship with the main character. We

emphasize main and supporting characters by thickening the stroke of

the nodes, applying diverging colors, and including labels; their colors

can be changed using a color picker.

Panel Layout. Since each panel contains a narrative, its size and

placement are crucial for conveying the message. Speciﬁcally, viewers

can infer event duration from the width of the panel [64] and determine

the chronological order based on the panel’s position in the strip, where

time proceeds from left to right and top to bottom [50, 64]. DG Comics

is designed to generate an appropriate panel layout. The layout algo-

rithm ﬁrst calculates the number of rows, or tiers, by taking the square

root of the total panel number speciﬁed by the user, ensuring each panel

maintains a minimum height. For example, with 16 panels, DG Comics

assigns 4 tiers. It then allocates panels to each tier to balance the sum

of their time intervals. The comic is partitioned using this structure,

with gutters between panels for clarity. This approach normalizes time

points across tiers, assuming that panels with more time points convey

more information, and encodes event duration into panel size [4].

Captions. DG Comics provides a caption template for each panel in

the main character’s point of view to assist storytelling. Each caption

consists of three clauses about a summary, major change, and the most

inﬂuential relationship. The summary describes expansions, contrac-

tions, and constancy based on the difference between the graphs before

and after. It includes a prepositional phrase in front to indicate the

timespan and reference time for a change. Then we compare the total

number of nodes added, deleted, and preserved, and state one case that

is the maximum. Lastly, we select the node(s) that have the highest

link weights and mark whether the main character had obtained, lost,

strengthened, or weakened its relationship with those nodes. A set of

clauses is created per main character, but each clause can be combined

into one with multiple subjects if different main characters share the

same content. When the panel portrays a particular time point, it simply

reports the strongest relationship.

5.2.2 Interaction Design

DG Comics offers diverse editing interactions to enhance ﬂexibility

in content creation (R5). Users can manage the general format of the

canvas and graph drawing using the toolbar (Fig. 3C, top-right) in the

Graph Comic View and adjust the details, such as text and graph styles

within the canvas using the toolbox that appears on the right side of

Graph Comic View when users interact with the element.

Canvas Editing. We deﬁne an editable component of the panel as a

canvas. DG Comics includes three canvas types: graph, text, and image.

An auto-generated panel contains one graph and two text canvases—a

caption and temporal information. Users can move, resize, and rotate

a canvas by interacting with the four sides and corners of the canvas,

respectively. The canvas editing can be done using the toolbar on the

upper right corner of the Graph Comic View (Fig. 3C). Clicking the

button reveals a list of canvas types that can be added to the workspace.

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

978

Hanspeter P ster 0 0.38 0.06

J. Edward Swan II

0 0.37 0.12

Daniel A. Keim

0 0.37 0.03

David H. Laidlaw

0 0.36 0.03

Xiaoru Yuan

0 0.36 0.03

W. Brent Seales

0 0.35 0.18

Hangzai Luo

0 0.35 0.18

Christy Jie Liang

0 0.35 0.09

Summary Node Attribute Community

eigenvector_centrality

Search node

Node Attribute Distance Min. Max Avg

Hanspeter P ster 0 0.38 0.06

J. Edward Swan II

0 0.37 0.12

Daniel A. Keim

0 0.37 0.03

David H. Laidlaw

0 0.36 0.03

Xiaoru Yuan

0 0.36 0.03

W. Brent Seales

0 0.35 0.18

Hangzai Luo

0 0.35 0.18

Christy Jie Liang

0 0.35 0.09

1 2 3 4 5

eigenvector_centrality

Search node

Node Attribute Distance Min. Max Avg

Fig. 5: Node Attribute Table. This table shows a list of all the nodes

and their values over time.

One distinct option is

Background

, which inserts a background image

into the graph canvas. To add a graph canvas, users select the timespan

on the Summary View by brushing before pressing the

Graph

button.

This will generate the change summary of an extracted main character

during the chosen timespan as explained in Sec. 4.1. The

Text

option

enables users to create an empty text box, while the

Image

option offers

preset icons such as speech balloons and arrows and supports importing

external images via the button. Users can simply delete, and move

back and forth the focused canvas by clicking , , and buttons.

Graph Drawing. DG Comics provides multiple modes for generat-

ing the graph layout before manually repositioning. When pressing the

button on the toolbar, buttons for three modes appear. The default

mode is a

Basic

force-directed graph that uses the concepts of repulsion

and attraction to place the relevant nodes closer and irrelevant nodes

further. Derived from this graph structure, the

Filled

mode reduces

empty space to a minimum in the way of multiplying scale to the rel-

ative position. This mode can be useful to increase the readability of

nodes by preventing overlapping. The

Fixed

mode supports mental

map preservation by ﬁxing node positions across the panels (Fig. 3M).

To achieve this, we compute a force-directed layout of the union of all

nodes and links in the dynamic graph over its full timespan. This is

useful for tracking changes in individual panels and facilitates the easy

detection of nodes and links being added or deleted.

Text Editing. When the text canvas gets focused, it triggers a

movable toolbox to appear on the upper right side of the Graph Comic

View. Users can drag the text content for selection and change it to be

bold, italic, or underlined. The typeface and size are also adjustable.

The background and border of the text canvas can be on and off. We

integrate an LLM (

Mixtral 8x7B

[36]); clicking the button will

use the LLM to improve the caption automatically.

Graph Element Editing. In the default mode, users can zoom in

on the graph canvas by scrolling and pan by dragging the background.

Each node is movable, and its style, along with the link, is editable.

DG Comics offers three methods for selecting multiple nodes: users

can individually click on the nodes; employ the lasso tool from the

graph comic view toolbar for easier group selection; or use a supporting

character table by clicking on a character row and then the

EDIT

button

(Fig. 3E). Once the selection is made, group movement of the nodes

is available, and a toolbox appears, allowing customization of styles.

The node-level toolbox provides functions for changing color, size,

and stroke properties, adjusting opacity, applying highlights, adding

labels, inserting images, deleting nodes, and clustering nodes. Links

can be edited by a direct click, and the link-level toolbox offers similar

functions, except for labeling, grouping, and image insertion. It also

features the ability to change markers, particularly useful for directed

graphs. Users can change the color and opacity of a cluster created

either by manual grouping or by selecting from the Community View

(Sec. 5.5). Once all the editing is done, users can export the ﬁnal

product as PNG, JPG, or PDF via the button.

5.3 Timeline View

We can express temporal changes not only with the symbolic repre-

sentation but also through the separation of panels. The Timeline View

(Fig. 3F) lets users choose a transition style by listing static graphs of

different time points. Users can scrutinize graphs included within the

chosen time span and select one to add to the canvas (R4, R5).

The Timeline View provides two functions for generating a series

of graphs: adding and replacing the selected panel. The

ADD

option

enables switching focus from overview to detail by attaching the graphs

of certain points in the bottom right corner. We use the same layout

computation for the

ADD

option to embed multiple panels in a limited

space while not covering the root panel completely. The

REPL

option

replaces the selected panel and arranges new panels in temporal order.

5.4 Tables

We provide three tables in DG Comics to select node attributes and

characters to display in the graph comic. The common features of the

tables are sorting by values and searching by names.

The Node Attribute Table (Fig. 5) facilitates node exploration and

assists in manual main character selection by providing comprehensive

attribute values and chart thumbnails (R3). Users can switch to this

table by clicking the second tab above the Summary View (Fig. 3G).

The line chart thumbnail tracks the value over time; lines are dashed if

data is missing (Fig. 5B). To prevent possible bias from inconsistent

scales, we also mark minimum, maximum, and average values. Users

can examine different attributes of a node—e.g., the total number of

links, PageRank, and centrality—by selecting the option in the top

left corner of the view (Fig. 5A). The heatmap presents the degree of

dissimilarity between the consecutive time points, meaning that the

color gets darker as the distance computed with our modiﬁed Jaccard

index increases (Fig. 5C). Using this thumbnail, users can capture

moments with signiﬁcant change. When selecting a main character on

the attribute table, the dendrogram is re-created to reﬂect the subgraph-

level change (R2). This means that each time node of the dendrogram

represents the subgraph grounded in the selected node (Fig. 3A).

The character tables (Fig. 3D, E) manage the nodes depicted in the

graph comic. The list of nodes is updated whenever users click the

graph canvas to focus. Users can either add or delete nodes by selecting

or deselecting the leftmost checkbox. While the Main Character Table

(Fig. 3D) shows all the nodes as candidates for the main character, the

Supporting Character Table (Fig. 3E) lists the neighbor nodes that are

linked with the currently selected main character (R3). The minimum,

maximum, and average values in the Main Character Table reﬂect the

attribute selected in the Node Attribute Table. The total weight of links

to a node in the selected time span is an indication of the importance

of the node (Fig. 3E). The Class column in the Main Character Table

enables direct manipulation of graph element styles. By entering the

tag name as input and clicking on the Class header, users can open

the CSS editing view to redeﬁne the properties of the corresponding

nodes (Fig. 3D). The

EDIT

button in the Supporting Character Table is

explained in Section 5.2.2.

5.5 Community View

Identifying the communities in which nodes are involved (i.e., commu-

nity membership) and tracking their evolution is crucial for a deeper

understanding of their relationships (R2). For instance, if two nodes

from different communities merge into the same community, we can

assume that their relationships become stronger; if they split, the rela-

tionship becomes weaker. A Sankey diagram is a suitable approach to

illustrate such ﬂow by representing communities as nodes and transi-

tions between them as links. To differentiate their nodes and links from

graph elements, we call them community nodes and paths. However,

its readability signiﬁcantly decreases with the numerous overlapping

paths and varied sizes of community nodes. Given that the narrative of

the graph comic is character-driven, we decided to highlight only the

communities of the chosen characters in this work. Moreover, we re-

designed the Sankey diagram to enable fair inspection and comparison

of community changes associated with the characters. Fig. 6 shows our

Community View, re-designed in this work.

In a traditional Sankey diagram, the size of a community node is

proportional to the size of the community, making it difﬁcult to observe

transitions for characters in smaller communities. To overcome this

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

979

kim ET AL.: DG COmiCS: SEmi-AUTOmATiCALLY AUTHORiNG GRAPH COmiCS FOR DYNAmiC GRAPHS

Summary Node Attribute Community

Louvain

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

RESET SELECTION RESET PATH

ain

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

Fig. 6: Community View changes when communities are selected,

as shown in the left inset, where red and blue colors mean Keim and

Schreck, respectively, whose communities diverged since 2014.

obstacle, we encode the community size into a sequential color scheme

while normalizing the community node size. Users can choose a cluster

color in a color palette by selecting the second checkbox on the ‘Sankey’

column (Fig. 3 D, E). We set the time as the x-axis and place community

nodes from top to bottom based on their community size. To connect

the community nodes for a selected character, we adopt a circular

shape with padding inside the grid cell so that the path can enclose it

smoothly. As a result, the Community View displays the colored path of

community evolution for the selected character while blurring nodes

outside of this path as Fig. 6 shows.

The Community View enables users to detect six types of events

in community evolution (Fig. 7): growth, contraction, merging, split-

ting, birth, and death [56]. Since DG Comics supports character-based

exploration for generating narratives, the community is primarily iden-

tiﬁed by the character. Therefore, the start and end of the path indicate

the birth and death of the character’s community within the dataset’s

time span. Users can determine growth and contraction through color

changes and the vertical position of the circle across different years.

Furthermore, by selecting multiple characters, users can observe how

distinct communities evolve to merge and split at speciﬁc time points

throughout the time span. Notably, the visual patterns depicted by

multiple paths can illustrate how the communities evolved through

various events (e.g., splitting, merging, homogeneity, heterogeneity)

over different time periods as Fig. 7 shows. After exploring community

evolution, users can add a cluster to the canvas by clicking the commu-

nity node (Fig. 3O). DG Comics uses Bubble Sets [21] to highlight the

group of characters involved. The

RESET PATH

button (upper left) clears

all highlights from the Community View, while the

RESET SELECTION

button

removes all the clusters from the graph comic. The default setting dis-

plays the communities detected by the Louvain algorithm, but another

option is available in the drop-down list, such as afﬁliations of authors

in the case of Vispubdata.

XAMPLE

We demonstrate the use of DG Comics in scientometric analysis with

data from the visualization community. We use Vispubdata [35], a

dataset containing information on 3,620 papers from IEEE Visualiza-

tion conferences (InfoVis, SciVis, VAST, and VIS) from 1990 to 2022,

including conference, year, title, paper type, authors, afﬁliation, key-

words, and number of citations. This dataset is frequently used for

cross-checking in dynamic network research and is highly relevant to

the visualization community. To create graph comics with compelling

storylines, we constructed author networks, where nodes represent

authors and links represent collaborations based on co-authored papers.

6.1 Analysis

We start with the default dendrogram, which computes the dissimilarity-

based hierarchy of the entire dynamic graph, adjusting the depth slider

to segment the time into nine divisions. This process generates nine

panels, offering a visual narrative of the data, which in this case, is

the collaboration network in VIS papers. We select 1.5-level ego

networks to observe connections between the neighbor nodes as well.

Birth

Death

Contraction

Growth

Splitting

Merging

Heterogeneity

Homogeneity

Fig. 7: Community evolution. Six archetypes of community evolution.

The auto-generated comic reveals DANIEL A. KEIM as a central ﬁgure

in two distinct periods, 2006–2011 and 2012–2016, with the network

expanding in the earlier period and contracting in the latter. We can now

generate a graph comic from Keim’s perspective for closer examination.

By opening the Node Attribute Table and searching for “Keim,” we

detect the high dissimilarity (Fig. 5C) and peak in eigenvector central-

ity (Fig. 5B) between 2011 and 2012, which suggests transformations

in Keim’s collaborative patterns. With this insight, we create a new

dendrogram for Keim and regenerate the graph comic, which unveils

two main clusters before and after 2011, indicating major events in the

network’s evolution (Fig. 3A). We move the sliders to ﬁlter out impact-

ful relationships involving the top 15 percent of total collaborations and

highlight the top 5 percent (Fig. 3B). To examine changes closely, we

add the graph canvas by brushing 2011 to 2020 from the Summary View

and insert the individual graphs of 2011 and 2012 from the Timeline

View (Fig. 3F). The tool’s features allow us to preserve the mental map

(Fig. 3M), helping us to capture the establishment of new relationships

and community shifts. Notably, we observe a strengthening of rela-

tionship with T

OBIAS

CHRECK

from 2011 to 2020, prompting us to

further compare the dynamics between Keim and Schreck.

In the Community View, we color-code Keim’s (blue) and Schreck’s

(red) communities and draw paths to visualize their changes over time

(Fig. 6). Despite multiple collaborations, their communities have di-

verged since 2014 (Fig. 3O), reﬂecting the impact of Schreck moving

to Graz University of Technology in 2015. This divergence, despite

ongoing collaboration, illustrates nuanced dynamics common in aca-

demic networks. Further exploration reveals D

OMINIK

ACHA

as a

signiﬁcant yet eventually diverging connection, mirroring the evolution

of community afﬁliations and the impact of career milestones such as

Ph.D. completion and industry employment. By juxtaposing Sacha’s

trajectory with Schreck’s, we enrich the narrative, offering compara-

tive insights into how individual careers and collaborations shape the

broader academic community. This scenario demonstrates the tool’s ca-

pacity to dissect, visualize, and narrate complex relational data, making

it an invaluable resource for understanding dynamic networks.

VALUATION

To explore the usefulness and effectiveness of DG Comics, we con-

ducted a comprehensive user study. This study includes usability tests

and interviews to gather valuable insights from participants and in-

depth feedback from experts to validate its versatile application.

7.1 User Study

We conducted a controlled user study to evaluate DG Comics. The

study consisted of a pre-study questionnaire, a tutorial session, graph

comics creation, and a post-questionnaire with interviews. We present

both quantitative ratings and qualitative insights from the interviews.

We used the COW trade dataset [9,10] of international trades from 1870

to 2014, covering 207 unique nations. Using this data, we built a di-

rected graph of international trade where each node and link represents

nationality and corresponding trade (in U.S. dollars), respectively.

7.1.1 Participants

We recruited 13 participants from a university (4 females) aged from

23 to 29 (

M = 25.4, SD = 2.2

). The number of participants is consis-

tent with prior work [17, 34, 54], aligning with recommendations for

usability testing. They were graduate (

n = 8

) and undergraduate (

n = 5

)

students from electrical engineering (

n = 4

), computer science (

n = 5

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

980

2 1 10

2 7 4

3 9 1

21 6 4

62 4 1

321 3 4

7 6

321 4 5

strongly

disagree

strongly

agree

Q1) I was able to eﬃciently create graph comics using the system.

Q2) I could eﬀectively express the story I wanted using the system.

Q3) Using the system was enjoyable.

Q4) The system helped in discovering new insights or perspectives from the data.

Q5) The graph comics created with the system were easy to understand.

Q6) I think the graph comics were represented reasonably.

Q7) I think the graph comics were aesthetically presented.

Q8) I am satisﬁed with the outcome.

Overall Ratings of DG Comics

4.46

(0.66)

4.46

(0.52)

4.00

(0.91)

4.62

(0.77)

4.15

(0.69)

3.85

(0.55)

3.31

(0.85)

3.54

(1.33)

IQR Median

20 40 60 80 1000

67.88 (7.76)

System Usability Scale

Fig. 8: DG Comics ratings. Overall ratings (top) and SUS score (bottom).

Values are means and standard deviations (in parentheses).

and artiﬁcial intelligence (

n = 4

). All participants had experience in

analyzing data using Python and visualizing it with Python, Excel, and

PowerPoint. Most of them had experience dealing with graph data (e.g.,

trafﬁc, sensor network, etc.), except for three participants who were

aware of graph data but lacked experience in processing it. Participants

were compensated $22 (USD) for 2 hours of study. The study was

reviewed and deemed exempt by our Institutional Review Board.

7.1.2 Procedure

The study began with a pre-study questionnaire that included: (1)

demographic information, such as age, gender, and education level;

(2) participant experience with data analysis, data visualization, and

graph data; and (3) participant familiarity with graph comics, the trade

dataset, and events of world history relevant to the study. We then

explained what a data comic is and how it can be applied to graph

data. We conducted a 20-minute tutorial on how to use DG Comics and

interpret the visual representations. Then participants were given 5 to

10 minutes to explore the system on their own and ask questions about

its functions. Note that we used Vispubdata [35] for the tutorial session

to prevent learning effects on the dataset used in the experiment.

After participants felt comfortable with the tool, we introduced them

to an international trade dataset spanning from 1930 to 1960. They

were tasked with using DG Comics to craft an engaging graph comic,

highlighting the evolving relationships between nations. To aid in their

task, we provided access to a Wikipedia page titled Timelines of Modern

History, which includes sections such as Timeline of the 20th Century

and List of Wars. The purpose was to accommodate varying levels of

familiarity with world history among participants. While participants

could use search engines to gather information and craft their storylines,

we prohibited the use of any comprehensive historical resources, such

as videos, that might present a ﬁnished interpretation of events.

Assistance was offered solely at the request of participants, ensuring

that their interaction with the system remained independent. Partici-

pants used Google Chrome in full-screen on two 32-inch monitors with

a resolution of

2560 × 1600

pixels—one for information search and the

other for the actual comic creation. Participants were allotted 1 hour

and 10 minutes for this task, during which their interactions with the

system were screen-recorded for further analysis.

Following the comic creation phase, we asked participants to ﬁll out

a post-study questionnaire. This questionnaire gauged their opinions

on the usability, overall experience, and the helpfulness of different

views within DG Comics, as well as their satisfaction with their ﬁnal

product, using a 5-point Likert scale. We concluded each session with a

semi-structured interview to explore the reasoning behind questionnaire

responses, the narratives constructed, and comprehensive feedback on

the system’s functionality and user experience.

7.1.3 Quantitative Results

We adopted a 5-point Likert scale to assess the overall experience of

the system (Q1-Q4) and the output that participants generated (Q5-

Q8). We used the System Usability Scale (SUS) designed by Brook

et al. [15] for global evaluation of systems usability. Interpreting the

Likert scales as ordinal, we report the quantitative results in a discrete

visualization style [67], the histogram [32, 84] with medians (as blue

lines) and quartiles (in gray areas). We also provide means and standard

deviations considering intervalist views [19].

As Fig. 8 shows, participants felt positive about their experience

using the system (Q1-Q3). In particular, they highly valued discov-

ering new insights and perspectives from the system (Q4). They all

agreed that the comics they made were easy to understand (Q5-Q6).

However, aesthetics and satisfaction with the output varied among the

participants, likely due to the discrepancy between their editing skills

and expectations toward their outcomes (Q7-Q8). The SUS score is

acceptable (

M = 67.88, SD = 7.76

); compare this to the average of 68.2

from 273 prior studies [8].

7.1.4 Qualitative Results

We analyzed participants’ responses and video recordings associated

with the scores. Here we present these qualitative results.

The summary view supports efﬁcient story generation. Most

participants reported that the overview with the dendrogram helped

generate stories (

MD = 4, M = 4.15, SD = 0.99

). Even though most

of them had little knowledge of the data, they were able to understand

the summary of changes in relationships at a glance. They used the

auto-generated comic as a starting point. The hierarchical clustering

effectively grouped similar time points to depict relevant stories into

one panel, allowing participants to “choose how detailed or brief to

express the story by adjusting the level (P5).” Clusters gave a hint to

users where to focus; for example, P9 after detecting groups before and

after World War II, “focused on the speciﬁc branches for storytelling

the events” while P12 “scrutinized the clusters as a chunk of changes.”

The community view effectively aids in inferring relationships

between nodes and events. Participants used the community view

to check the relationship between nodes (

n = 7

). Some participants

displayed a node’s community to see other nodes in the same cluster

and used this information to structure the graph (P3, P7, P8, P11,

P12). P12 excluded some countries such as the U.S. from the graph

since “their consistency in the relationship with the United Kingdom

suggests no special event happening.” They instead captured insights by

inspecting the community visualization. For example, the comparison

between the community evolution paths of different countries provides

an implication for a signiﬁcant event when there is a split or a merge.

P6 searched the interaction between Poland and Germany since “their

paths overlapped during the 1930s.”

Participants wanted to use the system with their own data. All

participants indicated that the system helped them effectively express

the story (

MD = 4, M = 4.46, SD = 0.52

). Some commented on chal-

lenges in visualizing dynamic graphs, especially for big and complex

data (P1, P5). However, using DG Comics, they were able to convey

changes in the relationship between nodes over time regardless of fa-

miliarity with the data (

n = 7

). P7 noted that it became much easier to

narrate the story of network evolution because “the system helps with

the hardest part, [which is] selecting the main character.” P10 said that

although he did “not have an aesthetic sense, [he liked that he could]

make a satisfactory product within an hour.” Some participants brought

up the point that the convenience of the system does not only rely on

visualization but also on interpretation of the data. “If the series of time

points are clustered, users can think about the reason and assume the

circumstance (P5).” P1, who conducts research on tracking individuals

using ultra-wideband data, said, “it would be very helpful to [be able

to] detect when interference between two groups occurs.”

7.2 Expert Feedback

We conducted online interviews with six experts across various do-

mains, who were invited via emails with consideration of domains and

expertise. Each expert has a Ph.D. degree and 8 to 19 years of work

experience in their respective ﬁelds. Their areas of expertise include

Communication (E1: 19 years, E2: 8 years), History (E3: 12 years),

Design (E4: 12 years), Management Information System (E5: 8 years),

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

981

kim ET AL.: DG COmiCS: SEmi-AUTOmATiCALLY AUTHORiNG GRAPH COmiCS FOR DYNAmiC GRAPHS

and Economics and Finance (E6: 11 years). Before interviews, we sent

them the information on the interviews (e.g., goal, duration, expected

outcomes), a demo video, and a website link to the deployed system.

The interviews began with a short system demonstration, followed by a

Q&A session for any clariﬁcations. Once interviewees were familiar

with the system, we proceeded to discuss its various aspects.

All appreciated the usefulness of the tool in telling a story using an

approachable format (i.e., comic). They reported various strengths of

DG Comics, including automated story extraction, rich functionality

in editing for story presentation, and novelty of the approach in DG

Comics. E1 lauded “the system is capable of extracting outstanding

stories with the corresponding timespan, streamlining the research

process for data journalists.” E4 and E5 also noted that laypeople,

such as marketers or end-user creators, can use the tool to create and

present a story with data that would otherwise have been inaccessible

to them. E1 and E2 raised a similar point, that DG Comics enriches

users’ insights by providing the change in dynamics between characters.

E3 even expressed his willingness to incorporate the system into his

curriculum of political history. “Given the possible impact of politics

on economics, I believe students can gain insights into economic re-

lations, such as international trade, by comparing them with political

knowledge through a comic." E5 and E6 underlined the contribution

of autogenerated output to efﬁcient decision-making. Speciﬁcally, E5

commented that autogenerated output from dynamic graphs will help

users in management who need to analyze dynamically changing re-

lationships among companies—“an intuitive understanding of supply

ﬂows and the impacts of subcontractor changes can help users make

strategic decisions in supply chain management.” E4 noted that many

people need to create and present stories from data but are often frus-

trated by challenges brought by new and complex data. In this situation,

he expected the tool would be very helpful. He highlighted the novelty

of the approach that DG Comics provides, stating, “... I give 10 out of

10 for novelty because I have not seen this type of system and support

from a complex data storytelling perspective.”

All experts estimated that the target user spectrum is broad, encom-

passing a diverse array of individuals. With the automatic generation

of comic templates and robust editing functions for crafting user-driven

stories, experts believe DG Comics meets the needs of a wide range

of users, from novices to professionals. They also commented that

usage scenarios with DG Comics may vary depending on the user’s

level of expertise and goals, even among the same expert group. For

instance, E1 predicted that “novice users would primarily utilize the

Graph Comic View, whereas experts would navigate between views

to gain a deeper understanding of the data.” Additionally, E4 men-

tioned that “researchers and data analysts would review the summary

of changes from a generated comic to select data before beginning their

analysis, yet if they already know the key points, their focus would shift

toward authoring to convey their message.”

Some experts expressed concerns about the steep learning curve due

to the comprehensive features implemented in the system. For example,

E2 and E4 expressed a concern that this system seems to require some

learning time before users fully enjoy the system functions. They

suggested creating a lite version of the system with fewer features for

beginners. On the other hand, some other experts, namely E2, E5,

and E6, indicated that experts may need additional features that help

users automatically process unstructured, complex graph datasets. E5,

in particular, mentioned that there is a signiﬁcant amount of dynamic

graph data in companies, but only data scientists can pre-process the

data for analysis and presentation purposes—“If the tool included an

interface for processing raw dynamic graph data, it could have a more

substantial impact, given the number of employees in enterprises who

need to analyze, monitor, and present stories with large graph datasets.”

8LIMITATIONS AND FUTURE WORK

DG Comics provides a ﬂexible and intuitive authoring environment for

graph comics creation. However, the numerous functions and steps in

the authoring process may cause newcomers to face a steep learning

curve. To mitigate this, providing an interactive tutorial and a help

menu that supports keyword search and mouse-over explanations can

be a reasonable solution. We also consider providing a function for

adjusting the level of interface complexity. Since our user study may

not capture the full diversity of all potential users and was conducted in

a limited amount of time, further investigation, such as a longitudinal

study, would help ensure ecological validity in the complexity reduction

of visualization and storytelling tools across different users.

The wide range of domain experts we interviewed suggests that

DG Comics has potential for diverse applications. However, ques-

tions about data preprocessing remain. Expert feedback reveals that

real-world network data, such as transactions between companies, com-

munications within and between departments, and distribution channels,

are often unstructured and unprocessed. To address this, automated

preprocessing of raw data into a compatible format is the next crucial

step for DG Comics to enhance its versatility. As a prototype, DG

Comics does not provide automated methods for extracting interesting

patterns from node and link attributes (e.g., anomaly detection for time

series [66]). Future work may focus on developing automated methods

to improve the utility and generalizability of storytelling tools. We also

expect the tool can be extended to incorporate datasets from biology

(DNA, proteins), software engineering (call graphs, code dependencies,

developer activities), and road networks with varying trafﬁc situations.

While the dendrogram effectively encodes each snapshot as a leaf

node, it faces scalability limitations with large dynamic networks con-

taining numerous snapshots. To enhance readability, we need to explore

abstraction techniques [44, 52, 57] such as collapsing and expanding

snapshots, despite the additional interaction they introduce. Currently,

the prototype supports only a linear cut on the dendrogram. Though

users can create additional canvases by brushing, enabling multiple

cuts at different abstraction levels would offer greater ﬂexibility in gen-

erating templates automatically. A future system supporting automated

story generation with multiple cuts at different levels would improve

both scalability and usability for real-world applications.

Besides, examining more sophisticated ﬁltering techniques and com-

munity discovery methods would signiﬁcantly enhance functionality.

DG Comics offers sliders for intuitive ﬁltering, allowing users to high-

light important nodes based on link weights and reduce visual clut-

ter. However, adopting more advanced ﬁltering techniques should be

inspected to prevent potential biases [60] from selecting subgraphs

centered on the main character. The Community View displays commu-

nities detected by the Louvain algorithm or predeﬁned communities

such as afﬁliation. To enrich narratives, future work could incorporate

advanced community discovery methods for temporal networks [61],

enabling the tool to suggest more intriguing and comprehensive stories.

9CONCLUSION

This paper presents DG Comics, a semi-automatic authoring tool for

graph comics based on identifying notable changes in a dynamic

graph. Our approach automatically extracts meaningful events us-

ing a causality-preserving hierarchical clustering method. With the

algorithmic output, users can adjust granularities by manipulating the

aggregation level and reﬁne aesthetics by making editorial decisions

down to minute details in comic strips through an intuitive user interface.

Our user evaluation and expert feedback demonstrate the usefulness of

the system, providing promising areas for future work.

CKNOWLEDGMENTS

This work was supported by the National Research Foundation of Korea

(NRF) grant funded by the Korea government (MSIT) (No.RS-2023-

00218913, No. 2021R1A2C1004542), by a grant of the Korea Health

Technology R&D Project through the Korea Health Industry Develop-

ment Institute (KHIDI), funded by the Ministry of Health & Welfare,

Republic of Korea (grant number:HI22C0646), and by the Institute of

Information & Communications Technology Planning & Evaluation

(IITP) grants (No. 2024-00360227-Leading Generative AI Human

Resources Development, No. 2020-0-01336–Artiﬁcial Intelligence

Graduate School Program, UNIST). Niklas Elmqvist was funded by

Villum Investigator grant VL-54492 by Villum Fonden.

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 31, NO. 1, JANUARY 2025

982

REFERENCES

[1]

J.-w. Ahn, C. Plaisant, and B. Shneiderman. A task taxonomy for network

evolution analysis. IEEE Transactions on Visualization and Computer

Graphics, 20(3):365–376, 2014. doi: 10.1109/TVCG.2013.238 2, 3

[2]

D. Archambault, T. Munzner, and D. Auber. Grouseﬂocks: Steerable

exploration of graph hierarchy space. IEEE Transactions on Visualization

and Computer Graphics, 14(4):900–913, 2008. doi: 10.1109/TVCG.2008.

34 2

[3]

B. Bach, N. Henry-Riche, T. Dwyer, T. Madhyastha, J.-D. Fekete, and

T. Grabowski. Small MultiPiles: Piling time to explore temporal patterns

in dynamic networks. Computer Graphics Forum, 34(3):31–40, 2015. doi:

10.1111/cgf.12615 2, 3

[4]

B. Bach, N. Kerracher, K. W. Hall, S. Carpendale, J. Kennedy, and N. H.

Riche. Telling stories about dynamic networks with Graph Comics. In

Proceedings of the ACM Conference on Human Factors in Computing

Systems, pp. 3670–3682. ACM, New York, NY, USA, 2016. doi: 10.

1145/2858036.2858387 1, 2, 3, 5

[5]

B. Bach, E. Pietriga, and J.-D. Fekete. Visualizing dynamic networks with

matrix cubes. In Proceedings of the ACM Conference on Human Factors

in Computing Systems, pp. 877–886. ACM, New York, NY, USA, 2014.

doi: 10.1145/2556288.2557010 2

[6]

B. Bach, N. H. Riche, S. Carpendale, and H. Pﬁster. The emerging genre

of data comics. IEEE Computer Graphics and Applications, 37(3):6–13,

2017. doi: 10.1109/MCG.2017.33 2, 3

[7]

B. Bach, Z. Wang, M. Farinella, D. Murray-Rust, and N. H. Riche. Design

patterns for data comics. In Proceedings of the ACM Conference on

Human Factors in Computing Systems, pp. 1–12. ACM, New York, NY,

USA, 2018. doi: 10.1145/3173574.3173612 2, 3

[8]

A. Bangor, P. Kortum, and J. Miller. Determining what individual sus

scores mean: Adding an adjective rating scale. Journal of usability studies,

4(3):114–123, 2009. 8

[9]

K. Barbieri, O. M. Keshk, and B. M. Pollins. Trading data: Evaluating our

assumptions and coding rules. Conﬂict Management and Peace Science,

26(5), 2009. 7

[10]

K. Barbieri and O. M. G. Keshk. Correlates of war project trade dataset

codebook, version 4.0, 2016. https://correlatesofwar.org/. 7

[11]

F. Beck, M. Burch, S. Diehl, and D. Weiskopf. A taxonomy and survey of

dynamic graph visualization. Computer Graphics Forum, 36(1):133–159,

2017. doi: 10.1111/cgf.12791 2, 3

[12]

M. Behrisch, B. Bach, N. Henry Riche, T. Schreck, and J.-D. Fekete.

Matrix reordering methods for table and network visualization. Computer

Graphics Forum, 35(3):693–716, 2016. doi: 10.1111/cgf.12935 2

[13]

A. Bezerianos, P. Dragicevic, J. Fekete, J. Bae, and B. Watson. Ge-

neaQuilts: A system for exploring large genealogies. IEEE Transactions

on Visualization and Computer Graphics, 16(6):1073–1081, 2010. doi: 10

.1109/TVCG.2010.159 2

[14]

M. Bostock, V. Ogievetsky, and J. Heer. D

: Data-driven documents. IEEE

Transactions on Visualization and Computer Graphics, 17(12):2301–2309,

2011. 4

[15]

J. Brooke et al. Sus: A quick and dirty usability scale. Usability evaluation

in industry, 189(194):4–7, 1996. 8

[16]

M. Burch, M. Hlawatsch, and D. Weiskopf. Visualizing a sequence of a

thousand graphs (or even more). Computer Graphics Forum, 36(3):261–

271, 2017. doi: 10.1111/cgf.13185 2

[17]

K. Caine. Local standards for sample size at chi. In Proceedings of the

ACM Conference on Human Factors in Computing Systems, pp. 981–992.

ACM, New York, NY, USA, 2016. 7

[18]

E. Cakmak, U. Schlegel, D. Jäckle, D. Keim, and T. Schreck. Multiscale

snapshots: Visual analysis of temporal summaries in dynamic graphs.

IEEE Transactions on Visualization and Computer Graphics, 27(2):517–

527, 2020. doi: 10.1109/TVCG.2020.3030398 2, 3

[19]

J. Cariﬁo and R. Perla. Resolving the 50-year debate around using and

misusing likert scales. Medical education, 42(12):1150–1152, 2008. 8

[20]

Q. Chen, S. Cao, J. Wang, and N. Cao. How does automation shape the

process of narrative visualization: A survey on tools. IEEE Transactions

on Visualization and Computer Graphics, 30(8):4429–4448, 2024. doi: 10

.1109/TVCG.2023.3261320 2

[21]

C. Collins, G. Penn, and S. Carpendale. Bubble sets: Revealing set

relations with isocontours over existing visualizations. IEEE Transactions

on Visualization and Computer Graphics, 15(6):1009–1016, 2009. doi: 10

.1109/TVCG.2009.122 7

[22]

A. Dal Col, P. Valdivia, F. Petronetto, F. Dias, C. T. Silva, and L. G. Nonato.

Wavelet-based visual analysis of dynamic networks. IEEE Transactions

on Visualization and Computer Graphics, 24(8):2456–2469, 2017. doi: 10

.1109/TVCG.2017.2746080 2, 3

[23]

C. Dunne and B. Shneiderman. Motif simpliﬁcation: improving network

visualization readability with fan, connector, and clique glyphs. In Pro-

ceedings of the ACM Conference on Human Factors in Computing Systems,

pp. 3247–3256. ACM, New York, NY, USA, 2013. doi: 10.1145/2470654.

2466444 2

[24]

N. Elmqvist and J. Fekete. Hierarchical aggregation for information visual-

ization: Overview, techniques, and design guidelines. IEEE Transactions

on Visualization and Computer Graphics, 16(3):439–454, 2010. doi: 10.

1109/TVCG.2009.84 2, 3

[25]

V. Filipov, A. Arleo, and S. Miksch. Are we there yet? A roadmap

of network visualization from surveys to task taxonomies. Computer

Graphics Forum, 42(6):e14794, 2023. doi: 10.1111/cgf.14794 2, 3

[26]

X. Gao, B. Xiao, D. Tao, and X. Li. A survey of graph edit distance.

Pattern Analysis and Applications, 13, 2010. 4

[27]

T. Gartner, P. Flach, S. Wrobel, B. Scholkopf, and M. Warmuth. On graph

kernels: Hardness results and efﬁcient alternatives. In Proceedings of

Learning Theory and Kernel Machines, vol. 2777, pp. 129–143, 2003. 4

[28]

S. Ghani, N. H. Riche, and N. Elmqvist. Dynamic insets for context-aware

graph navigation. Computer Graphics Forum, 30(3):861–870, 2011. doi:

10.1111/j.1467-8659.2011.01935.x 2

[29]

S. Hadlak, H. Schumann, C. H. Cap, and T. Wollenberg. Supporting the

visual analysis of dynamic networks by clustering associated temporal

attributes. IEEE Transactions on Visualization and Computer Graphics,

19(12):2267–2276, 2013. doi: 10.1109/TVCG.2013.198 2

[30]

N. Henry and J.-D. Fekete. MatrixExplorer: a dual-representation system

to explore social networks. IEEE Transactions on Visualization and

Computer Graphics, 12(5):677–684, 2006. doi: 10.1109/TVCG.2006.160

[31]

N. Henry, J.-D. Fekete, and M. J. McGufﬁn. NodeTrix: a hybrid visualiza-

tion of social networks. IEEE Transactions on Visualization and Computer

Graphics, 13(6):1302–1309, 2007. doi: 10.1109/TVCG.2007.70582 2

[32]

J. Huang, A. Mishra, B. C. Kwon, and C. Bryan. Conceptexplainer: Inter-

active explanation for deep neural networks from a concept perspective.

IEEE Transactions on Visualization and Computer Graphics, 29(1):831–

841, 2022. doi: 10.1109/TVCG.2022.3209384 8

[33]

J. Hullman, S. M. Drucker, N. H. Riche, B. Lee, D. Fisher, and E. Adar.

A deeper understanding of sequence in narrative visualization. IEEE

Transactions on Visualization and Computer Graphics, 19(12):2406–2415,

2013. doi: 10.1109/TVCG.2013.119 3

[34]

W. Hwang and G. Salvendy. Number of people required for usability

evaluation: the 10±2 rule. Comm. of the ACM, 53(5):130–133, 2010. 7

[35]

P. Isenberg, F. Heimerl, S. Koch, T. Isenberg, P. Xu, C. Stolper, M. Sedl-

mair, J. Chen, T. Möller, and J. Stasko. vispubdata.org: A Metadata Col-

lection about IEEE Visualization (VIS) Publications. IEEE Transactions

on Visualization and Computer Graphics, 23, 2017. doi: 10.1109/TVCG.

2016.2615308 7, 8

[36]

A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bamford,

D. S. Chaplot, D. de las Casas, E. B. Hanna, F. Bressand, G. Lengyel,

G. Bour, G. Lample, L. R. Lavaud, L. Saulnier, M.-A. Lachaux, P. Stock,

S. Subramanian, S. Yang, S. Antoniak, T. L. Scao, T. Gervet, T. Lavril,

T. Wang, T. Lacroix, and W. E. Sayed. Mixtral of experts. CoRR,

abs/2401.04088, 2024. doi: 10.48550/ARXIV.2401.04088 6

[37]

B. Kale, M. Sun, and M. E. Papka. The state of the art in visualizing

dynamic multivariate networks. Computer Graphics Forum, 42(3):471–

490, 2023. doi: 10.1111/cgf.14856 3

[38]

D. Kang, T. Ho, N. Marquardt, B. Mutlu, and A. Bianchi. ToonNote:

Improving communication in computational notebooks using interactive

data comics. In Proceedings of the ACM Conference on Human Factors

in Computing Systems, pp. 1–14. ACM, New York, NY, USA, 2021. doi:

10.1145/3411764.3445434 2, 3

[39]

H. Kashima, K. Tsuda, and A. Inokuchi. Marginalized kernels between

labeled graphs. In Proceedings of the International Conference on Machine

Learning, p. 321–328, 2003. 4

[40]

N. W. Kim, N. H. Riche, B. Bach, G. Xu, M. Brehmer, K. Hinckley,

M. Pahud, H. Xia, M. J. McGufﬁn, and H. Pﬁster. DataToon: Drawing

dynamic network comics with pen + touch interaction. In Proceedings of

the ACM Conference on Human Factors in Computing Systems, pp. 1–12.

ACM, New York, NY, USA, 2019. doi: 10.1145/3290605.3300335 2, 3

[41]

P. M. Kogge. Jaccard coefﬁcients as a potential graph benchmark. In

Proceedings of IEEE International Parallel and Distributed Processing

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.

983

kim ET AL.: DG COmiCS: SEmi-AUTOmATiCALLY AUTHORiNG GRAPH COmiCS FOR DYNAmiC GRAPHS

Symposium Workshops, pp. 921–928, 2016. 4

[42]

R. Kosara and J. Mackinlay. Storytelling: The next step for visualization.

IEEE Computer, 46(5):44–50, 2013. doi: 10.1109/MC.2013.36 3

[43]

N. M. Kriege, P.-L. Giscard, and R. C. Wilson. On valid optimal assign-

ment kernels and applications to graph classiﬁcation. In Advances in

Neural Information Processing Systems, p. 1623–1631, 2016. 4

[44]

B. Lee, C. S. Parr, C. Plaisant, B. B. Bederson, V. D. Veksler, W. D.

Gray, and C. Kotﬁla. Treeplus: Interactive exploration of networks with

enhanced tree layouts. IEEE Transactions on Visualization and Computer

Graphics, 12(6):1414–1426, 2006. doi: 10.1109/TVCG.2006.106 9

[45]

B. Lee, N. H. Riche, P. Isenberg, and S. Carpendale. More than telling

a story: A closer look at the process of transforming data into visually

shared stories. IEEE Computer Graphics and Applications, 35(5):84–90,

2015. doi: 10.1109/MCG.2015.99 3

[46]

W. Li, S. Schöttler, J. Scott-Brown, Y. Wang, S. Chen, H. Qu, and B. Bach.

Networknarratives: Data tours for visual network exploration and analysis.

In Proceedings of the ACM Conference on Human Factors in Computing

Systems, CHI ’23. Association for Computing Machinery, New York, NY,

USA, 2023. doi: 10.1145/3544548.3581452 2

[47]

G. Ma, N. K. Ahmed, T. L. Willke, and P. S. Yu. Deep graph similarity

learning: a survey. Data Mining and Knowledge Discovery, 35(3):688–

725, 2021. doi: 10.1007/S10618-020-00733-5 4

[48]

P. Mahe and J.-P. Vert. Graph kernels based on tree patterns for molecules.

Machine Learning, 75:3–35, 2009. 4

[49]

T. May, M. Steiger, J. Davey, and J. Kohlhammer. Using signposts for

navigation in large graphs. Computer Graphics Forum, 31(3pt2):985–994,

2012. doi: 10.1111/j.1467-8659.2012.03091.x 2

[50]

S. McCloud. Understanding Comics: The Invisible Art. William Morrow

Paperbacks, 1994. 2, 3, 4, 5

[51]

F. McGee, M. Ghoniem, G. Melançon, B. Otjacques, and B. Pinaud. The

state of the art in multilayer network visualization. Computer Graphics

Forum, 38(6):125–149, 2019. doi: 10.1111/cgf.13610 2, 3

[52]

K. Misue. Area-adaptive drawing of rooted trees. In Proceedings of the

IEEE Paciﬁc Symposium on Visualization, pp. 152–161, 2024. 9

[53]

A. Narayanan, M. Chandramohan, R. Venkatesan, L. Chen, Y. Liu, and

S. Jaiswal. graph2vec: Learning distributed representations of graphs.

CoRR, abs/1707.05005, 2017. 2

[54] J. Nielsen. Usability engineering. Morgan Kaufmann, 1994. 7

[55]

C. Nobre, M. Meyer, M. Streit, and A. Lex. The state of the art in

visualizing multivariate networks. Computer Graphics Forum, 38(3):807–

832, 2019. doi: 10.1111/cgf.13728 2

[56]

G. Palla, A.-L. Barabási, and T. Vicsek. Quantifying social group evolution.

Nature, 446(7136):664–667, 2007. doi: 10.1038/nature05670 7

[57]

C. Plaisant, J. Grosjean, and B. B. Bederson. Spacetree: Supporting explo-

ration in large node link tree, design evolution and empirical evaluation.

In Proceedings of the IEEE Symposium on Information Visualization, pp.

57–64, 2002. doi: 10.1109/INFVIS.2002.1173148 9

[58]

J. Ramon and T. Gartner. Expressivity versus efﬁciency of graph kernels.

In Proceedings of the First International Workshop on Mining Graphs,

Trees and Sequences, pp. 65–74, 2003. 4

[59]

N. H. Riche, C. Hurter, N. Diakopoulos, and S. Carpendale. Data-Driven

Storytelling. A K Peters/CRC Press, Boca Raton, FL, USA, 2018. 2

[60]

L. E. Rocha, N. Masuda, and P. Holme. Sampling of temporal networks:

Methods and biases. Physical Review E, 96(5):052302, 2017. 9

[61]

G. Rossetti and R. Cazabet. Community discovery in dynamic networks:

a survey. ACM Computing Surveys (CSUR), 51(2):1–37, 2018. 9

[62]

M. Ruži

cka. Anwendung mathematisch–statisticher methoden in der

geobotanik (synthetische bearbeitung von aufnahmen), 1958. 4

[63]

B. Saket, P. Simonetto, and S. G. Kobourov. Group-level graph visualiza-

tion taxonomy. CoRR, abs/1403.7421, 2014. doi: 10.48550/arXiv.1403.

7421 2

[64] M. Saraceni. The Language of Comics. Psychology Press, 2003. 4, 5

[65]

E. Segel and J. Heer. Narrative visualization: Telling stories with data.

IEEE Transactions on Visualization and Computer Graphics, 16(6):1139–

1148, 2010. doi: 10.1109/TVCG.2010.179 1, 2, 3

[66]

M. Shin, J. Kim, Y. Han, L. Xie, M. Whitelaw, B. C. Kwon, S. Ko, and

N. Elmqvist. Roslingiﬁer: Semi-automated storytelling for animated

scatterplots. IEEE Transactions on Visualization and Computer Graphics,

29(6):2980–2995, 2023. doi: 10.1109/TVCG.2022.3146329 9

[67]

L. South, D. Saffo, O. Vitek, C. Dunne, and M. A. Borkin. Effective use of

likert scales in visualization evaluations: A systematic review. Computer

Graphics Forum, 41(3):43–55, 2022. doi: 10.1111/cgf.14521 8

[68]

C. D. Stolper, B. Lee, N. H. Riche, and J. Stasko. Emerging and recurring

data-driven storytelling techniques: Analysis of a curated collection of

recent stories. Technical Report MSR-TR-2016-14, Microsoft Research,

2016. 3

[69]

S. Suh, J. Zhao, and E. Law. CodeToon: Story ideation, auto comic gener-

ation, and structure mapping for code-driven storytelling. In Proceedings

of the ACM Symposium on User Interface Software and Technology, pp.

13:1–13:16. ACM, New York, NY, USA, 2022. doi: 10.1145/3526113.

3545617 2, 3

[70]

C. Tong, R. C. Roberts, R. Borgo, S. P. Walton, R. S. Laramee, K. Wegba,

A. Lu, Y. Wang, H. Qu, Q. Luo, and X. Ma. Storytelling and visual-

ization: An extended survey. Information, 9(3):65, 2018. doi: 10.3390/

info9030065 2, 3

[71]

S. van den Elzen, D. Holten, J. Blaas, and J. J. van Wijk. Dynamic network

visualization with extended massive sequence views. IEEE Transactions

on Visualization and Computer Graphics, 20(8):1087–1099, 2014. doi: 10

.1109/TVCG.2013.263 2

[72]

S. van den Elzen, D. Holten, J. Blaas, and J. J. van Wijk. Reducing

snapshots to points: A visual analytics approach to dynamic network

exploration. IEEE Transactions on Visualization and Computer Graphics,

22(1):1–10, 2015. doi: 10.1109/TVCG.2015.2468078 2, 3

[73]

S. Varma, S. Shivam, A. Thumu, A. Bhushanam, and D. Sarkar. Jaccard

based similarity index in graphs: A multi-hop approach. In Proceedings

of IEEE Delhi Section Conference, pp. 1–4, 2022. 4

[74]

C. Vehlow, F. Beck, and D. Weiskopf. Visualizing group structures in

graphs: A survey. Computer Graphics Forum, 36(6):201–225, 2017. doi:

10.1111/cgf.12872 2, 3

[75]

T. Von Landesberger, A. Kuijper, T. Schreck, J. Kohlhammer, J. J. van

Wijk, J.-D. Fekete, and D. W. Fellner. Visual analysis of large graphs:

state-of-the-art and future research challenges. Computer Graphics Forum,

30(6):1719–1749, 2011. doi: 10.1111/j.1467-8659.2011 .01898.x 2, 3

[76]

Q. Wang, Z. Li, S. Fu, W. Cui, and H. Qu. Narvis: Authoring narrative

slideshows for introducing data visualization designs. IEEE Transactions

on Visualization and Computer Graphics, 25(1):779–788, 2019. doi: 10.

1109/TVCG.2018.2865232 2

[77]

Z. Wang, J. Ritchie, J. Zhou, F. Chevalier, and B. Bach. Data comics for

reporting controlled user studies in human-computer interaction. IEEE

Transactions on Visualization and Computer Graphics, 27(2):967–977,

2021. doi: 10.1109/TVCG.2020.3030433 2

[78]

Z. Wang, H. Romat, F. Chevalier, N. H. Riche, D. Murray-Rust, and

B. Bach. Interactive data comics. IEEE Transactions on Visualization

and Computer Graphics, 28(1):944–954, 2022. doi: 10.1109/TVCG.2021.

3114849 2, 3

[79]

Z. Wang, S. Wang, M. Farinella, D. Murray-Rust, N. H. Riche, and

B. Bach. Comparing effectiveness and engagement of data comics and

infographics. In Proceedings of the ACM Conference on Human Factors

in Computing Systems, pp. 1–12. ACM, New York, NY, USA, 2019. doi:

10.1145/3290605.3300483 2

[80]

F. Xia, K. Sun, S. Yu, A. Aziz, L. Wan, S. Pan, and H. Liu. Graph learning:

A survey. IEEE Transactions on Artiﬁcial Intelligence, 2(2):109–127,

2021. doi: 10.1109/TAI.2021.3076021 2

[81]

J. Xu, Y. Tao, Y. Yan, and H. Lin. Exploring evolution of dynamic networks

via diachronic node embeddings. IEEE Transactions on Visualization and

Computer Graphics, 26(7):2387–2402, 2018. doi: 10.1109/TVCG.2018.

2887230 2, 3

[82]

J. Yan, X.-C. Yin, W. Lin, C. Deng, H. Zha, and X. Yang. A short survey

of recent advances in graph matching. In Proceedings of the International

Conference on Multimedia Retrieval, p. 167–174, 2016. 4

[83]

V. Yoghourdjian, T. Dwyer, K. Klein, K. Marriott, and M. Wybrow. Graph

thumbnails: Identifying and comparing multiple graphs at a glance. IEEE

Transactions on Visualization and Computer Graphics, 24(12):3081–3095,

2018. doi: 10.1109/TVCG.2018.2790961 2

[84]

J. Zhao, S. Xu, S. K. Chandrasegaran, C. Bryan, F. Du, A. Mishra, X. Qian,

Y. Li, and K. Ma. ChartStory: Automated partitioning, layout, and cap-

tioning of charts into comic-style narratives. IEEE Transactions on Vi-

sualization and Computer Graphics, 29(2):1384–1399, 2023. doi: 10.

1109/TVCG.2021.3114211 3, 8

[85]

Z. Zhao, W. Benjamin, N. Elmqvist, and K. Ramani. Sketcholution: Inter-

action histories for sketching. International Journal of Human-Computer

Studies, 82:11–20, 2015. doi: 10.1016/j.ijhcs.2015.04.003 3

[86]

Z. Zhao, R. Marr, and N. Elmqvist. Data comics: Sequential art for data-

driven storytelling. Technical Report 15, Human-Computer Interaction

Laboratory, University of Maryland, College Park, 2015. 1, 2, 3

Authorized licensed use limited to: INRIA. Downloaded on December 11,2025 at 13:39:36 UTC from IEEE Xplore. Restrictions apply.